Can AI run a vending operation? New study raises doubts

Four of five artificial intelligence agents managing a simulated vending network generated lower financial returns than a conventional inventory management system, according to a new study that raises questions about how much operational control businesses should give autonomous AI.

The study, "AutoVend: Benchmarking LLM Agents on Long-Horizon Autonomous Vending Operations," was conducted by researchers affiliated with StepFun, Zhejiang University, Xi'an Jiaotong-Liverpool University and other institutions. The researchers evaluated five AI agents managing 20 vending machines over 180 simulated days, using an environment calibrated with actual commercial vending sales data.

The agents sometimes made decisions that appeared beneficial individually but ultimately undermined financial performance through excessive inventory purchases, unnecessary service visits and poor coordination of operating expenses.

Industry experts say the results do not mean AI has no place in vending. Instead, they suggest that combining AI capabilities with established inventory management software and human oversight may be more practical than turning over entire operations to autonomous agents.

Testing AI against real-world vending challenges

The researchers developed a simulation involving 20 vending machines across factories, college campuses, residential properties, commercial establishments and offices.

Each AI agent managed the network for 180 simulated days, beginning with approximately $6,000 in operating capital. Responsibilities included selecting products, setting prices, purchasing inventory and scheduling replenishment visits.

The simulation incorporated a shared warehouse, delivery delays, product expiration, storage expenses and labor costs. Researchers calibrated the environment using six months of transaction records from 99 machines operated by an unidentified commercial smart vending company. The data covered 220,128 products sold from a catalog of 1,372 products.

Five large language models were evaluated against a conventional inventory management system that used historical demand, product availability and established replenishment calculations.

Only GPT-6.1 Sol consistently outperformed the conventional system, finishing with average net assets of approximately $46,400, compared with $31,600 for the rule-based system. That represents approximately 47% more ending net assets than the conventional benchmark.

The other four models finished below the benchmark, although all five generated positive net asset growth during the simulation.

The results illustrate both the potential and limitations of autonomous AI in vending operations. While one model substantially exceeded conventional software, most struggled to coordinate the interconnected financial and logistical decisions involved in managing a vending network.

Manuel Mojica, founder of Vending Circle and owner of MM Healthy Vending, said AI's effectiveness depends heavily on the quality and completeness of the information available about individual machines.

"If a vending machine location has too many unknown variables, the AI will struggle to make decisions on that machine," Mojica said in written comments to Vending Times.

He said operators using AI-powered inventory monitoring still need the ability to intervene when equipment or operating conditions require attention.

[RELATED MEDIA]

Where AI decisions go wrong

One of the study's central findings was that strong performance in individual operating categories did not necessarily translate into better overall financial results.

Some agents minimized product waste but struggled to keep machines stocked. Others maintained product availability while accumulating excessive inventory or incurring unnecessary replenishment expenses.

For example, a DeepSeek V4.1 Flash agent scheduled consecutive visits to a campus vending machine that generated approximately 77 cents in gross profit between visits, even though each service visit cost about $5.25.

Another agent repeatedly purchased inventory despite having more than 2,000 unsold units of one product in the warehouse.

The researchers also found substantial differences in how models balanced waste and availability. Claude Sonnet 5.5 recorded an inventory waste rate of just 0.16% but experienced completely empty shelves on more than 25% of operating site-days.

By comparison, DeepSeek V4.1 Flash experienced completely empty shelves on just 1.67% of operating site-days but recorded a higher inventory waste rate of 1.85%.

These examples demonstrate how decisions that appear reasonable in isolation can undermine financial performance when their cumulative effects are considered. Minimizing spoilage, for example, offers little benefit if customers routinely encounter empty machines.

Mojica said operators need systems that recognize when automated decisions should be reviewed by a person.

"There should be a break point in the system that forces the human to take action if it reaches a certain price or amount of stock," he said.

A hybrid approach to vending management

Chris Selland, founder and CEO of AI research firm Differential Factor and a lecturer at Northeastern University's D'Amore-McKim School of Business, said the findings should not be interpreted as evidence that large language models are unsuitable for vending operations.

"What it's really showing is that they're the wrong tool for the day-to-day job of running the operation," Selland said in an email interview.

Replenishment, pricing and scheduling are established operational problems that conventional inventory management systems already handle effectively, he said.

AI may offer greater value in less structured tasks, such as interpreting supplier communications, identifying unusual activity and explaining operational exceptions.

"Let the proven technology continue working on the everyday, repetitive work - and deploy AI where there's more variance and uncertainty," Selland said.

Christopher Graney-Ward, CEO of ToonSoft and a former member of Mitsubishi HC Capital UK's AI governance board, agreed that operators should distinguish between structured calculations and tasks requiring interpretation.

Numerical calculations, inventory constraints and scheduling decisions can be handled by specialized software, while AI agents analyze unstructured information and determine when to use those tools.

However, Graney-Ward said in an email interview that the leading AI model's performance also demonstrates that autonomous systems could outperform conventional software under certain conditions.

"The most important thing in your statistic isn't that four didn't surpass traditional software, but that one did - it's proof that it can be done, if the setup is right," he said.

That distinction is important for vending operators evaluating AI products. The study does not establish that autonomous AI is inherently inferior to conventional software. Rather, it suggests that performance depends on how systems coordinate inventory, purchasing, replenishment and financial decisions.

Why human oversight still matters

Industry experts emphasized that AI systems making financial decisions need clearly defined operating limits.

Graney-Ward recommended safeguards including purchasing limits, price boundaries, human approval requirements and the ability to reverse decisions.

"Agentic systems live and die by their controls," he said.

Hana Klimczak, chief operating officer and founder of Scaalr, described in an email interview how her company applies similar restrictions to AI agents handling maintenance requests for property management companies.

The agents can receive complaints, assess urgency and contact personnel responsible for emergencies. However, employees retain authority over decisions involving expenditures.

"We didn't want our AI agents to make any decisions that involve spending money," Klimczak said.

She suggested vending operators could adopt a comparable approach, allowing AI to identify sales patterns and inventory problems while established software performs calculations and employees approve consequential purchasing decisions. For vending businesses, such safeguards could prevent automated systems from placing excessive inventory orders, changing prices outside established limits or scheduling service visits that cost more than the revenue they are likely to generate.

Proving AI's value in commercial operations

The researchers also examined whether AI agents improved their operating strategies as they accumulated sales data.

Although some models improved individual performance measures, all five showed lower average daily asset growth during a later evaluation period than during an earlier one. The researchers cautioned that changing operating conditions prevented them from determining whether the agents' use of feedback caused those results.

The study also identified differences in AI processing costs. GPT-6.1 Sol generated the highest overall financial returns, while DeepSeek V4.1 Flash produced the greatest net asset growth per dollar of estimated AI processing expense.

For operators, those findings suggest that evaluating AI requires considering not only sales and inventory performance but also the recurring cost of operating the technology.

Graney-Ward recommended testing AI alongside existing management software before giving it direct control over commercial operations.

"To compare, run the agent alongside the existing system, starting in shadow mode where it recommends and a human acts, and measure margin, stockouts, spoilage and staff hours over a fixed period," he said.

Mojica similarly recommended comparing AI-managed machines with conventional systems at similar types of locations and measuring outcomes such as product waste.

For operators evaluating AI-powered vending management platforms, the findings suggest focusing on four operational considerations:

• Replenishment costs: Does the system account for labor, transportation and service expenses before scheduling machine visits?
• Product availability: Can it maintain acceptable stockout rates without generating excessive inventory?
• Inventory waste: Does it balance spoilage prevention with customer demand and warehouse stock levels?
• Financial controls: Are purchasing limits, pricing boundaries and human approval requirements built into the system?

Operators also should determine whether an AI platform integrates with existing vending management software and whether its recommendations can be tested before the system receives authority to make purchases or schedule service calls.

What the study does — and does not — prove

The AutoVend study has limitations. Researchers conducted three runs per AI model in a single simulated operating environment. Although the simulation was calibrated using actual vending sales records, AI agents did not operate live vending routes.

Some assumptions, including pricing responses and operating costs, were established by the researchers rather than independently validated against field performance. The paper is also a preprint, and its findings do not establish that all commercial AI-powered vending systems underperform conventional software.

The researchers recommended further testing under different operating conditions, including systems that verify inventory availability before making decisions and retain lessons from earlier operational mistakes.

For vending operators, the findings suggest that AI's immediate value may lie in helping employees interpret information, identify problems and make better decisions rather than replacing established management systems entirely.

The performance of the leading model also leaves open the possibility that more capable autonomous systems, equipped with appropriate operational safeguards, could eventually outperform conventional approaches.

For now, the central question is not simply whether AI can manage vending operations, but which decisions it should make independently and which should remain under the control of established software and human operators.

As Selland put it, "Match the tool to the job."